Ask what you returned to, changed, rejected, or worked on over time.
Your history, made navigable
Find the questions, facts, and ideas you keep returning to.
Choose your ChatGPT export. Memory Map reads its conversation files in this tab, then builds a searchable semantic graph with a browser-loaded embedding model.
See all 51 things Memory Map can find →Need the ZIP?How to export your ChatGPT data
- Sign in to ChatGPT, open your profile menu, then choose Settings → Data controls.
- Under Export data, choose Export, then Confirm export.
- Open the email or SMS from OpenAI, download the ZIP, and bring it here without unzipping it.
Exports can take up to 7 days. The download link expires 24 hours after it arrives.
Settings exports require a signed-in eligible account and are not available for ChatGPT Business or Enterprise workspaces.
Use Claude Code or Codex?How to pack your agent sessions
- Install it, on macOS or Linux:
curl -fsSL https://chatgpt.significanthobbies.com/install.sh | sh - See what it finds, without writing anything:
memory-pack --dry-run --list - Write the archive into the current folder:
memory-pack - Drop the
memory-pack-<date>.zipit names here, without unzipping it.
Your prompts and the assistant’s replies are packed. Tool calls, tool output, reasoning traces, and file contents are left behind, which is why an archive stays small enough to analyse in a browser tab.
Both CLIs prune old sessions, so memory-pack also reads the prompt history they keep outside their transcripts. Sessions they have already deleted are rebuilt from the prompts that survive.
The installer downloads one small binary and checks it against a published checksum. memory-pack itself then runs entirely on your machine, makes no network calls, and masks credential-shaped tokens before it writes anything. On Windows, download the .exe from the releases page and run it directly.
Narrow it with --since 2026-01-01, or pack only sessions whose full transcript survives with --no-history.
- 1ParseRead split conversation files one at a time.
- 2MapEmbed bounded prompts and fact candidates.
- 3ExploreSearch, inspect repeats, and trace every result.
Private by construction.No archive upload endpoint, account, or default archive persistence. Model files download from Hugging Face; your conversation text is not sent with that request. Site-level page-view and performance telemetry is separate and receives no archive text or derived content.
Reading archive
Building your memory map
Inspecting ZIP entries…
Total elapsedUnder 1 sec
Estimated remainingEstimating…
Compute routeSelecting…
Ingestion sketch · deterministic routes
Every conversation becomes a stop.
- Open archive
- Read conversations
- Initial insights
- Prepare model
- Map semantic memory
- Assemble report
Statistics appear before embeddings finish. The first semantic run downloads and caches the selected model in your browser.
Route interrupted
This archive could not be mapped.
Mapped memory
Your conversation history, connected.
Analysis detailsSemantic work is running · balanced evidenceTiming, model, and confidence
Live execution route
How long your map took
Initial insights are ready. Semantic analysis is continuing.
- Initial insight
- —
- End to end
- Running…
- Model preparation
- Waiting
- Semantic work
- Waiting
- Compute route
- Selecting…
- Candidate coverage
- Forming…
- Open archive
- Read conversations
- Initial insights
- Prepare model
- Map semantic memory
- Assemble report
Filter inferred evidence
Analysis lens
Semantic candidates will be filtered without re-reading your ZIP.
Network scale
Overview
Seven ways to see the archive
Visual atlas
Applied to chronological charts, query tone, and language signals. What changed remains an all-history comparison.
Full parsed exportEmbedding-backed sampleOpen a mark to inspect sources
Read activity data
Read question-mix data
Wording signals, not feelings, personality, or diagnosis.
Read language-signal data
Read repeat data
Topic movement is still forming.The deterministic views are already usable.
Read topic-momentum data
Read conversation-shape data
Not enough months to compare yet.Three or more months of history are needed to call a shift.
Read what-changed data
Things connected by meaning
Topic map
Node size is conversation count. Lines are semantic proximity.
Semantic map is still forming.Deterministic insights are ready below.
Focused neighborhood · the complete topic list follows.
Ask the graph
Memory Chat
Your question traverses the map first. A small local model sees only the six retrieved evidence stops—not the full ZIP.
- Model
- LaMini-Flan-T5 77M
- Download
- ≈105 MB · cached by your browser
- Boundary
- Local q8 synthesis · English-focused
The chat model has not been loaded.
Visible retrieval route
How the graph answered
- Question received
- Searching semantic memory
- Traversing topic routes
- Building evidence pack
- Local synthesis
Why this memory architecture is different
The retrieval route, sources, changed and refuted statements, confidence controls, and local model boundary stay visible. Generated prose never becomes a new memory automatically.
This is a product-behavior prototype, not a claim about current ChatGPT internals. OpenAI’s own documentation says its legacy saved-memory experience could become stale or contradictory.
Read OpenAI’s Memory FAQ ↗Routes through time
How your map changed
Emerging, fading, and resurfacing are local trend labels—not explanations for why your interests changed.
Evolution routes appear after analysis.
Different ways to read the archive
Question lenses
Queries may appear in more than one domain. Open any result to inspect its supporting conversations.
Conservative writing signals
Likely typos
Conversations that changed tracks
Conversation strands
Interchange patterns
Questions you return to
Exact wording and semantic repeats are kept distinct.
Prompts from your own evidence
Questions your history asks you
Invitations to review a repeat, correction, old memory, quiet theme, or wording spike—not conclusions about you.
Wording, not personality
Query tone
A more expressive vocabulary
Language signals in your queries
What your sessions cost
Say less, do more
Issues, and what to do about them
Findings
Cost per turn by session length
Why long sessions compound
Every turn re-reads the session so far, so the same turn costs more the later it happens.
Where the spend concentrated
Your costliest sessions
Ranked by context re-read. A very long thread is usually several tasks that never got their own session.
Cadence over time
Rhythms
Selected total—
Peak month—
Monthly conversation starts across the selected period.
Read rhythm data
How the map was made
Transparent by default.
Statistics, question lenses, typo and thread-change candidates, query tone, language signals, and reflection prompts use local, inspectable rules. Topic relationships, semantic repeats, fact grouping, conversation strands, and search use normalized browser-loaded embeddings.
Assistant assertions are excluded from the fact ledger. Large histories are sampled for semantic work while totals continue to cover the full parsed export. Confidence controls refilter stored candidates; they do not make inferred evidence certain.